Document UC deductions review findings: floor scope, validated years, bias direction

[MaxGhenis]

Follow-up to #1815 addressing the methodology review's documentation and test items:

• Protected floor scope: notes on protected_floor.yaml that the modeled floor binds on all deduction categories, including last resort and child maintenance (current law allows those above the cap; JRF's briefing does not state whether their design exempts them — flagged for follow-up if it does).
• Validated years and regional composition: the latent rate distribution is national (regional factors scale incidence only) and validated against 2024–2025 observations; earlier years inherit it under their higher caps as an uncalibrated extrapolation.
• Direction of bias under cap reforms: lowering the cap further overstates newly constrained households and their gains; raising it rests on the distribution observed under the 25% cap.
• Annualization: annual per-household amounts are upper bounds for spell-limited deduction types (e.g. advance recovery).
• Aggregate caveat promoted to the changelog, with pointers to policyengine-uk-data#452 (UC caseload ~40% below admin counts — the binding limitation for floor-reform aggregates) and Add OpenFisca-UK methodology overview #450 (moving the imputation to dataset build).
• Test: zeroing all seven combination shares degrades gracefully (combination NONE, zero deductions, no crash) — covering the review's enum-fallback concern.

🤖 Generated with Claude Code